I am noticing that my ResNet model is showing some false positives in cases where novel negative example images are somewhat visually similar to positive examples.
In these cases, it's not simply that the probability of positive is greater than negative, but positive may be very high. This seems to be due to visual similarity in these images compared to the positive training examples, e.g. colors, orientations, etc.
I cannot discuss the subject matter and other technical details. The problem we're solving is not necessarily related to certain color patterns, but it is likely that negative examples may share colors and angles with the training images.
I'm looking to understand whether or not introducing negative examples of these kinds of things into my data set can help the model training to pay closer attention to other visual features that really matter as opposed to visual features that it may share with negative examples.